[Iccrg] updated RG CTCP review

[Lachlan Andrew]

Greetings Doug,

On 28/03/2008, Douglas Leith <Doug.Leith at nuim.ie> wrote:
> I'd suggest we wait for the revised draft to be circulated
>  before setting a short two week fuse on the review.
>  How do other people feel ?

I'd second that.

>  Re safety, personally I'd definitely like to see some tests carried
>  out over wireless before proceeding since its such a common last
>  hop.   The concern is that random access wireless (such as 802.11)
>  might have only a weak connection between delay and congestion, and
>  so have implications for the behaviour of delay algorithms.

Experiments would be good, but I'm not worried because
thought-experiments show it is safe.  Random fluctuations in delay
will reduce the delay-based component.  It might kill fairness, but
should not make it worse than a Reno flow, and we know that Reno isn't
clobbered too badly.

>  I also don't think I completely buy the low loss regime argument for
>  safety as it seems to be stated in the iccrg minutes.  That provides
>  safety in the sense of reverting to reno on lower speed links, but
>  surely we do need to think at least a little about safety on higher
>  speed links ?  If such links are rare so that we don't care about
>  safety, then we have no problem.  However, if they are not rare then
>  behaviour on those links should be considered, shouldn't it ?   Since
>  higher speed links here corresponds to a pipe size of 38 packets, I
>  suspect that pipes larger than this are common enough.

Yes, we need to think about it.  FWIW, my thoughts are that it is safe :)

It isn't the pipe size that decides if  dwnd  is used; it is the
Reno-cwnd.  If the Reno cwnd is larger than 38 packets, then the loss
rate is less than about (1/38^2).  If "safety" means "avoiding
congestion collapse", the question is whether congestion collapse can
happen with a low loss rate like that.

If we were pre-Tahoe and using Go-Back-N, then we could get congestion
collapse with an arbitrarily small loss rate if the window was large
enough.  However modern TCP uses Selective Retransmit (except after
timeouts, where the window is reduced to 1 MSS) which means that when
dwnd  is active we should get through at least about  (1 - 1/(38^2) )
of what we transmit.

(In case you're wondering:  No, I'm not on Murari's payroll...)

Perhaps we should have a more precise definition of what is meant by
"safe" and "congestion collapse".  If I recall correctly, the original
definition of congestion collapse was "adding an extra flow reduces
total network throughput".  That is clearly inadequate, because a
simple two-hop parking-lot is usually in congestion collapse according
to that definition: Adding an extra two-hop flow reduces the total
network throughput.

Cheers,
Lachlan


-- 
Lachlan Andrew  Dept of Computer Science, Caltech
1200 E California Blvd, Mail Code 256-80, Pasadena CA 91125, USA
Ph: +1 (626) 395-8820    Fax: +1 (626) 568-3603
http://netlab.caltech.edu/lachlan